Re: [PATCH 5/7] staging: usbip: Add encryption support to kernel

[Dominik Paulus]

Hi,

thank you very much for your feedback!

On Mon, Sep 23, 2013 at 12:59:29PM +0300, Dan Carpenter wrote:
> > +	while (total < size) {
> > +		uint32_t packetsize;
> > +		struct kvec recvvec;
> > +
> > +		/*
> > +		 * We use a global kfifo to buffer unrequested plaintext bytes.
> > +		 * Flush this buffer first before receiving new data.
> > +		 */
> > +		if (kfifo_len(&ud->recv_queue)) {
> > +			size_t next = min_t(size_t, kfifo_len(&ud->recv_queue),
> > +					size - total);
> > +			/* No error checking necessary - see previous line */
> > +			ret = kfifo_out(&ud->recv_queue, ((char *)
> > +						vec[0].iov_base)+total, next);
> 
> 
> The comment assume there is only one reader and one writer at a time,
> yes?  The casting is not needed:

Actually, we have not only one reader and one writer, but exactly one
thread doing all the work on the fifo. No parallel access is done at
all.

> 			ret = kfifo_out(&ud->recv_queue,
> 					vec[0].iov_base + total, next);
> 
> 
> v> +			total += next;
> > +			continue;
> > +		}
> > +
> > +		/* See usbip_sendmsg() for the format of one encrypted packet */
> > +
> > +		/*
> > +		 * Receive size of next crypto packet
> > +		 */
> > +		recvvec.iov_base = &packetsize;
> > +		recvvec.iov_len = sizeof(packetsize);
> > +
> > +		ret = kernel_recvmsg(ud->tcp_socket, msg, &recvvec, 1,
> > +				sizeof(packetsize), flags);
> > +		packetsize = be32_to_cpu(packetsize);
> > +		if (ret <= 0) {
> 
> I think a return of zero should mean total = -EBADMSG;.  In other words
> this check should be "if (ret < 0) {" and we hit the next else if.
> Same below again.

As we are wrapping kernel_recvmsg here, we wanted to leave the semantics
intact as far as possible. The calling code already checks for the correct
size.

> > +			total = ret;
> > +			goto err;
> > +		} else if (ret != sizeof(packetsize)) {
> > +			total = -EBADMSG;
> > +			goto err;
> > +		}

On Mon, Sep 23, 2013 at 01:35:04PM +0300, Dan Carpenter wrote:
> On Thu, Sep 19, 2013 at 04:11:57PM +0200, Dominik Paulus wrote:
> > +	if (crypto_aead_setkey(ud->tfm_send, sendkey, USBIP_KEYSIZE) != 0 ||
> > +			crypto_aead_setkey(ud->tfm_recv, recvkey,
> > +				USBIP_KEYSIZE) != 0 ||
> > +			crypto_aead_setauthsize(ud->tfm_send,
> > +				USBIP_AUTHSIZE) != 0 ||
> > +			crypto_aead_setauthsize(ud->tfm_recv,
> > +				USBIP_AUTHSIZE)) {
> > +		crypto_free_aead(ud->tfm_recv);
> > +		crypto_free_aead(ud->tfm_send);
> > +		kfifo_free(&ud->recv_queue);
> > +	}
> 
> This returns success on error instead of failure.

Indeed, that was horribly broken. Thanks.

On Mon, Sep 23, 2013 at 01:58:42PM +0300, Dan Carpenter wrote:
> On Thu, Sep 19, 2013 at 04:11:57PM +0200, Dominik Paulus wrote:
> > +{
> > +	struct crypto_aead *tfm;
> > +	struct aead_request *req;
> > +	struct tcrypt_result result;
> > +	struct scatterlist plain, cipher, assoc;
> > +	char iv[16];
> > +	u64 *iv_num;
> > +	u64 iv_net;
> > +	const int plainsize = packetsize - USBIP_AUTHSIZE;
> 
> Is it possible that packetsize is less than USBIP_AUTHSIZE?

No, currently, the caller (usbip_sendmsg() / usbip_recvmsg() are the
only functions calling usbip_crypt(), which itself is static) ensures
this.
Admittedly, this isn't great design. We added a check for packetsize <
USBIP_AUTHSIZE and an appropiate return here.

> > +	if (encrypt)
> > +		ret = crypto_aead_encrypt(req);
> > +	else
> > +		ret = crypto_aead_decrypt(req);
> > +
> 
> Good on you for figuring out what crypto_aead_en/decrypt() returns.
> Where are these functions documented?
> 
> > +        switch (ret) {
> > +        case 0: /* Success */
> > +                break;
> > +        case -EINPROGRESS:
> > +        case -EBUSY:
> > +                wait_for_completion(&result.completion);
> > +                break;
> > +        default:
> > +                aead_request_free(req);
> > +                return ret;
> > +        }
> > +

They aren't, actually. Documentation/crypto/api-intro.txt refers to the
regression test module, which uses exactly those return-values in
crypto/testmgr.c.
We noticed that wait_for_completion might not be the best idea, since it could
hang indefinitely, testmgr.c uses wait_for_completion_interruptible. Do we
want 'interruptible' or 'killable' here?
(Also we forgot to error-check result.err afterwards. We also fixed that.)

Regards,
Dominik Paulus, Tobias Polzer
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/